---
library_name: peft
base_model:
  - sensenova/SenseNova-SI-1.3-Qwen3-VL-8B
  - sensenova/SenseNova-SI-1.2-InternVL3-8B
  - Qwen/Qwen3-VL-8B-Instruct
pipeline_tag: video-text-to-text
tags:
  - lora
  - spatial-reasoning
  - vsi-bench
  - space-tokens
  - qwen3-vl
  - internvl3
---

# Chain-Spatial Models: Space Tokens checkpoints

These are the training checkpoints behind the *Space Tokens* paper, plus the later CVPR training runs. There are thirteen checkpoint directories, all **LoRA adapter checkpoints, plus auxiliary spatial-token weights**:

- **Original paper archive (LoRA rank 16 / alpha 32):** seven SenseNova checkpoints on top of [`sensenova/SenseNova-SI-1.3-Qwen3-VL-8B`](https://huggingface.co/sensenova/SenseNova-SI-1.3-Qwen3-VL-8B) (the final GRPO model, its SFT parent, the Stage-1-only lineage checkpoint, the VICA-only baseline and the three Table 4 token ablations), and the Table 1 Qwen + Space Tokens SFT model on top of [`Qwen/Qwen3-VL-8B-Instruct`](https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct).
- **CVPR training runs (LoRA rank 32 / alpha 64, added November 12, 2026):** InternVL Stage 0 / Stage 1 / GRPO on top of [`sensenova/SenseNova-SI-1.2-InternVL3-8B`](https://huggingface.co/sensenova/SenseNova-SI-1.2-InternVL3-8B) with a VGGT teacher, and Qwen Stage 0 / Stage 1 on top of `sensenova/SenseNova-SI-1.3-Qwen3-VL-8B` with a VGGT-Omega teacher.

The base models are not included. Each model's folder has its own `README.md` describing that model.

Training and evaluation code: [gitlab.com/hunterls97/covt](https://gitlab.com/hunterls97/covt).

This card is generated from [`checkpoints.MD`](checkpoints.MD), the catalogue that matches each result to a checkpoint and its evidence.

## Recent CVPR training backups — LoRA rank 32 / alpha 64

Five recent checkpoints were consolidated on **November 12, 2026**, in addition to the original paper archive below. Stages in this section follow the requested numbering: **0 = geometry alignment, 1 = supervised reasoning, 2 = GRPO**. GRPO itself uses code prompt stage 1 (supplied sparse spatial CoT).

| ID | Backbone / teacher | Stage and selection | Full VSI overall | Checkpoint |
|---|---|---|---:|---|
| `sn12_internvl_stage0_r32a64` | SenseNova-SI-1.2 InternVL3 / VGGT | Stage 0 — final, step 9851 | Not evaluated for selection | [SN12_InternVL3_SpaceTokens_stage0_VGGT_4gpu_eb32/2026-11-03_vggt_stage0_eb32_periodic_noeval/checkpoint-9851](SN12_InternVL3_SpaceTokens_stage0_VGGT_4gpu_eb32/2026-11-03_vggt_stage0_eb32_periodic_noeval/checkpoint-9851/) |
| `sn12_internvl_stage1_r32a64` | SenseNova-SI-1.2 InternVL3 / VGGT | Stage 1 — best, step 9000 | 69.284191 | [SN12_InternVL3_SpaceTokens_stage1_VGGT_4gpu_eb32/2026-11-05_from_original_stage0_9851/checkpoint-9000](SN12_InternVL3_SpaceTokens_stage1_VGGT_4gpu_eb32/2026-11-05_from_original_stage0_9851/checkpoint-9000/) |
| `sn12_internvl_grpo_r32a64` | SenseNova-SI-1.2 InternVL3 / VGGT | Stage 2 GRPO — original-prompt best, step 750 | 69.585845 | [SN12_InternVL3_SpaceTokens_GRPO_2gpu_eb8/2026-11-10_resume500_every250_182833/checkpoint-750](SN12_InternVL3_SpaceTokens_GRPO_2gpu_eb8/2026-11-10_resume500_every250_182833/checkpoint-750/) |
| `sn13_qwen_stage0_r32a64` | SenseNova-SI-1.3 Qwen3-VL / VGGT-Omega | Stage 0 — final, step 9851 | Not evaluated for selection | [SN13_SpaceTokens_stage1_8gpu_omega/2026-10-25_02-51-22/checkpoint-9851](SN13_SpaceTokens_stage1_8gpu_omega/2026-10-25_02-51-22/checkpoint-9851/) |
| `sn13_qwen_stage1_r32a64` | SenseNova-SI-1.3 Qwen3-VL / VGGT-Omega | Stage 1 — best retained full-split result, step 15000 | 67.046585 | [SN13_SpaceTokens_stage2_omega_epoch2/continued_run/checkpoint-15000](SN13_SpaceTokens_stage2_omega_epoch2/continued_run/checkpoint-15000/) |

**Selection and lineage:**

- **InternVL Stage 0:** the final checkpoint of `2026-11-03_vggt_stage0_eb32_periodic_noeval` completed epoch 1.0 at step 9851. This is the explicit parent of the selected reasoning run; the separately interrupted fixed-alignment experiment is not substituted for that parent.
- **InternVL Stage 1:** checkpoint 9000 ranks first among all ten saved reasoning checkpoints under both the historical evaluation and the updated scorer. Its historical full-VSI score was **69.626533**; rescoring and the subsequent GRPO initial evaluation give **69.284191**. Both are retained as different scoring provenance for the same model. This checkpoint is the frozen SFT reference/initial policy for GRPO.
- **InternVL Stage 2 / GRPO:** the user explicitly selected **checkpoint 750**, which leads the original native VSI evaluation at **69.585845**. The later evaluation with GRPO training-matched wording gives this checkpoint **68.381231** and selects checkpoint 500 instead (**68.386793**). These protocols are kept separate; checkpoint 500 is not substituted for the user's chosen backup.
- **Qwen Stage 0:** checkpoint 9851 is the completed one-epoch SenseNova-SI-1.3 alignment run at rank 32 / alpha 64. The original saved Trainer step is **9851**, while its DeepSpeed tag/global counter is **9850**; this existing discrepancy is preserved, not repaired during archival.
- **Qwen Stage 1:** checkpoint **15000** from the second supervised epoch has the best available full-split CoT score, **67.046585**, across the 14 retained full-VSI results from the original and continuation runs. The higher **70.597062** at step 6000 was measured on only **500** examples, and that checkpoint was pruned; it is not comparable to these 5,130-example scores. The two `stage2_*` Qwen run-directory names mean paper Stage 2 / code Stage 1 supervised training, **not GRPO**.
- **Qwen GRPO:** the user confirmed that no GRPO run was trained for this rank-32/alpha-64 model, so no such checkpoint is expected or marked missing.

The five **complete checkpoint directories** total **86.878 GB / 80.912 GiB** and **143 files**. Tokenizers, full auxiliary weights, saved model/phase configuration when present, Trainer state, RNG state and DeepSpeed optimizer shards are preserved. Original `.hydra/` folders remain beside the archived checkpoints; additional InternVL resume configurations and selection evidence are under [cvpr_r32a64_20261112](evidence/cvpr_r32a64_20261112/). These are adapter/auxiliary-state checkpoints; pretrained base weights and datasets remain separate. Original absolute paths in configs are preserved and may need remapping after relocation.

Every checkpoint file passed source/destination **SHA-256** and size verification, with independent copied files rather than hard links. Safetensors extents and ZIP member headers were checked without executing pickle payloads. No new inference or optimizer-resumption test was performed as part of copying. See [selection and verification](evidence/cvpr_r32a64_selection.json), [copy provenance](evidence/cvpr_r32a64_20261112/provenance.json), and the global integrity manifests below.

## Summary: accounted for and remaining experiments

| Paper table / experiment | Status | Details / remaining evidence needed |
|---|---|---|
| **Table 1 — SenseNova + Space Tokens** | Accounted for | Checkpoint 800 matches all reported scores. |
| **Table 1 — Qwen + Space Tokens** | Accounted for — accepted working identity | June 26 checkpoint 9851 copied as `qwen3_sft`, per user instruction. Fresh VSI score 64.0368 versus paper 64.1; numerical differences retained below. |
| **Table 1 — pretrained baselines** | Unverified | Model identities known; exact evaluation logs not found. |
| **Table 2 — OOD benchmarks** | Accounted for — experimenter confirmation | Stage 1+2 uses `sn13_sft` checkpoint 10000; corresponding Stage 1+2+3 OOD runs use `sn13_grpo` checkpoint 800. Original OOD logs/configs remain to be collected. |
| **Table 3 — VICA-only baseline** | Accounted for | Checkpoint 9851 matches 67.1. |
| **Table 3 — Stage 1 & 2** | Accounted for — identity confirmed | `sn13_sft` checkpoint 10000 is confirmed through the OOD experimenter. VSI logs show 68.667 versus paper 68.6; score discrepancy retained. |
| **Table 3 — Stage 1 & 2 & 3** | Accounted for | Same checkpoint as Table 1 SenseNova. |
| **Table 4 — all three token ablations** | Accounted for | Box-only, 3D-only, and combined checkpoints matched. |
| **Tables S1–S2 — hyperparameters** | No checkpoint needed | Configuration tables. |
| **Table S3 — reasoning-frame counts** | Partial | Six-frame SFT candidate preserved; need the four- and 32-frame evaluation logs/configs and confirmation of the six-frame result. |
| **Table S4 — token replacement** | Partial | Model identified; need exact full-split evaluation results and perturbation configs. |
| **Table S5 — attention masking** | Accounted for, with caveats | Checkpoint and both average scores matched; need clarification of backend/masking metadata and one 0.1-point task discrepancy. |
| **Additional — SenseNova Stage 1 only** | Archived — no paper table assigned | `sn13_stage1`, checkpoint 11000; parent explicitly selected for Stage 2. Preserved for lineage and future experiments. |

Entries marked Remaining, Tentative or Partial still need the original checkpoint path or run ID, with the matching evaluation log or config. Detailed checkpoint locations and supporting evidence follow below.

The original paper archive was consolidated October 20, 2026 against `review/space_tokens.pdf` (main paper and supplement). Eight original checkpoint directories were copied, totaling **168.985 GB / 157.380 GiB**, including tokenizer files, auxiliary weights, trainer state, RNG state, and DeepSpeed optimizer shards. Original files and paths inside configs were preserved, and the directories were uploaded here unchanged.

**Accounted for:** SenseNova Table 1 / Table 3 final model, the Stage 1+2 SFT model, Table 3 VICA-only baseline, all three Table 4 variants, and the checkpoint referenced by both Table S5 evaluations. The June 26 Qwen checkpoint is accepted as the working Table 1 identity by user instruction. The OOD experimenter confirmed the Stage 1+2 and Stage 1+2+3 checkpoint identities through the user. Numerical reproduction differences and missing original OOD result logs remain documented separately.

## Original paper checkpoint locations

Paths below are relative to the repository root. The seven `sn13_*` checkpoints use `sensenova/SenseNova-SI-1.3-Qwen3-VL-8B`; `qwen3_sft` uses `Qwen/Qwen3-VL-8B-Instruct`. These are **adapter checkpoints**, not complete copies of the pretrained base model.

| ID | Experiment / role | Paper VSI score | Logged full-split score | Checkpoint |
|---|---|---:|---:|---|
| `sn13_grpo` | Table 1 SenseNova + Space Tokens; Table 3 Stage 1 & 2 & 3; corresponding Stage 1+2+3 OOD runs; Table S5 model | 68.9 | 68.935723 | [SN13_SpaceTokens_GRPO_eval/2026-07-17_22-25-50/checkpoint-800](SN13_SpaceTokens_GRPO_eval/2026-07-17_22-25-50/checkpoint-800/) |
| `sn13_stage1` | Paper Stage 1 only (code stage 0); additional lineage checkpoint, no table result | — | Not assigned | [sensenova_1_3_st0_vsi500/2026-05-29_21-10-19/checkpoint-11000-universal](sensenova_1_3_st0_vsi500/2026-05-29_21-10-19/checkpoint-11000-universal/) |
| `sn13_sft` | Table 3 Stage 1 & 2; Table 2 OOD model (identity confirmed); likely Table S3 six-frame model | 68.6 | **68.667353** | [sensenova_1_3_st1_newprompt_vsi500/2026-06-23_21-17-15/checkpoint-10000](sensenova_1_3_st1_newprompt_vsi500/2026-06-23_21-17-15/checkpoint-10000/) |
| `sn13_vica` | Table 3 SenseNova + VICA-322K, no spatial task tokens | 67.1 | 67.070508 | [ST_ablation_SN13_VICA_only/2026-07-24_22-49-48/checkpoint-9851](ST_ablation_SN13_VICA_only/2026-07-24_22-49-48/checkpoint-9851/) |
| `sn13_box` | Table 4 box tokens only | 67.8 | 67.815140 | [ST_ablation_box_only/2026-07-27_19-57-56/checkpoint-2500](ST_ablation_box_only/2026-07-27_19-57-56/checkpoint-2500/) |
| `sn13_3d` | Table 4 3D reconstruction tokens only | 67.4 | 67.402361 | [ST_ablation_3D_only/2026-07-24_22-43-34/checkpoint-2500](ST_ablation_3D_only/2026-07-24_22-43-34/checkpoint-2500/) |
| `sn13_box_3d` | Table 4 box + 3D tokens, shorter ablation training | 68.4 | 68.416535 | [ST_ablation_box_3D/2026-07-27_21-34-19/checkpoint-2500](ST_ablation_box_3D/2026-07-27_21-34-19/checkpoint-2500/) |
| `qwen3_sft` | Table 1 Qwen + Space Tokens (accepted working identity) | 64.1 | **64.036810** | [clip_qwen3_st1/2026-06-26_16-50-09/checkpoint-9851](clip_qwen3_st1/2026-06-26_16-50-09/checkpoint-9851/) |

For the original numerical matches, archived evaluations explicitly identify the checkpoint and agree with the printed scores. OOD checkpoint identities are supported by experimenter confirmation relayed by the user. Qwen is an accepted working identification with a fresh full-split evaluation; its scores do not exactly reproduce the paper. The supplementary evaluations have additional provenance limitations below.

## Using the checkpoints

- These adapters need the project's own model wrappers, which add the spatial-token heads: `covt_qwen2_5_vl` for the Qwen3-VL-based checkpoints and the `SAT_InternVL3` wrapper for the InternVL3 checkpoints (named under `base_model_class` in each InternVL `adapter_config.json`). Both are part of the [covt project](https://gitlab.com/hunterls97/covt). A snapshot of the original loader and evaluation code is in [evidence/source_snapshot/](evidence/source_snapshot/). Plain `PeftModel.from_pretrained` will not restore the auxiliary weights in `non_lora_state_dict.bin`.
- Pass the base model explicitly. `base_model_name_or_path` in each `adapter_config.json` and the absolute paths in `.hydra/config.yaml` point to the original training server.
- The DeepSpeed shards under each `global_step*/` folder are only needed to resume training. They make up about 208 GiB of the repo (the rest is about 30 GiB), so for inference download with `--exclude "*/global_step*/*"`. One Qwen optimizer shard is corrupt in the original backup (see *Integrity verification*).
- Verify a download with `sha256sum -c SHA256SUMS`, run from the repository root.

## SenseNova Stage 1 only: additional lineage checkpoint

`sn13_stage1` preserves the original **paper Stage 1 geometry-alignment checkpoint (code stage 0)**. It is included for training lineage and future experiments; **no paper table result or full-split VSI score is assigned to it**.

- **Local checkpoint:** [sensenova_1_3_st0_vsi500/2026-05-29_21-10-19/checkpoint-11000-universal](sensenova_1_3_st0_vsi500/2026-05-29_21-10-19/checkpoint-11000-universal/).
- **Original source:** `/work/old_h200/moma/checkpoints/checkpoint-11000-universal`.
- **Identity:** `adapter_config.json` identifies SenseNova-SI-1.3-Qwen3-VL-8B with LoRA rank 16 / alpha 32. Strings inspected in `training_args.bin` identify run `sensenova_1_3_st0_vsi500` and its original May 29 run directory; that timestamp is used for the local catalogue path.
- **Selected training point:** global step **11000**, epoch **0.9771046612**, from a one-epoch run with planned maximum step **11258**. This is the checkpoint explicitly selected for Stage 2, rather than a claim that it is the final epoch-boundary checkpoint.
- **Stage 2 reference:** the preserved `sn13_sft` Hydra config names `/mnt/nfs/planning/shared/sat/checkpoints/sensenova_1_3_st0_vsi500/checkpoint-11000-universal` as `ckpt` with `load_tasks_only: true`. The archived loader imports the spatial task modules from this checkpoint and freezes them; this does not imply that it restores the complete Stage-1 LoRA adapter.
- **Contents and verification:** all **3,199 files / 40.672 GB** copied, including the adapter, auxiliary weights, tokenizer/processor, trainer/RNG state, original DeepSpeed shards, and universal-format state. Every source/destination SHA-256 matches; no structural archive issue was detected in this addition. No inference or training-resumption test was run.

The original Stage-1 `.hydra/` config was not present beside this exported checkpoint. Its identity is supported by saved checkpoint metadata and the child Stage-2 configuration; no missing configuration was reconstructed or invented. Supporting evidence, the child config, and the archived loader are preserved in [stage1_only_consolidation](evidence/stage1_only_consolidation/), including [provenance.json](evidence/stage1_only_consolidation/provenance.json).

## Alternative July Qwen checkpoint 3000 (not selected)

The other `covt` backup contains this approximately 2.7 GB adapter checkpoint:

`/work/old_h200/hunter/covt/train/outputs/qwen3-vl-SpaceTokens/2026-07-15_23-40-25/checkpoint-3000`

Its adapter config confirms **Qwen3-VL-8B-Instruct**, LoRA rank **16**, alpha **32**, and dropout **0.01**. The auxiliary weights include both 3D reconstruction and 3D-box heads. Its saved training output path links it to the matching July 15 run under `/work/old_h200/hunter/covt_grpo/train/outputs/`, whose Hydra config enables eight tokens per spatial task. These establish the model family and run lineage, but do not establish the paper's 64.1 result.

The binary W&B history for run **`0mvmdyz3`** was decoded directly because its final summary JSON is missing. It contains **3,001 history records through step 3000**, but evaluation metrics only at steps **0 and 1000**. The step-1000 scores are **49.8051** with the spatial prompt and **50.6777** with the plain prompt, on a smaller subset with only five task metrics; they cannot be compared directly with the paper's eight-task full-split 64.1 result. No evaluation explicitly referencing checkpoint 3000 was found in the searched logs.

This export has no `trainer_state.json` or model `config.json`, and all 738 adapter tensor names start with an extra `module.` prefix. Loading compatibility needs checking before a benchmark rerun. The original July checkpoint has not been modified or copied into this collection. The June 26 checkpoint below is now the accepted working Table 1 model; the July checkpoint is retained here as an alternative investigated during the search. Inspection details, decoded history, and related configs/logs are preserved in [qwen_checkpoint3000_audit](evidence/qwen_checkpoint3000_audit/).

## Qwen June 26 checkpoint 9851: accepted Table 1 model

`/work/old_h200/hunter/shared/clip_qwen3_st1/2026-06-26_16-50-09/checkpoint-9851`

A fresh evaluation completed on October 20, 2026 on GPUs 4–7, with **5,130 unique scored examples**, BF16, Flash Attention 2, batch size 1 per GPU, 32 frames, and the recorded paper image-resolution settings. All 40 auxiliary checkpoint tensors loaded with zero unexpected keys. The overall score is **64.0368096**, conventionally rounded to **64.0**, versus the paper's **64.1** (difference −0.0632 points). Its task vector does not match the printed row; relative direction is 52.6561 versus 55.1. The user has instructed us to treat this as the correct Table 1 Qwen checkpoint for now. Its full original checkpoint directory is consolidated locally as `qwen3_sft`; this working identification preserves the measured score differences and does not establish their cause. The original evaluations were on H100 hardware according to the user, while this rerun used H200; hardware has not been established as the reason for the difference.

| Metric | Paper | Fresh evaluation |
|---|---:|---:|
| Average | 64.1 | 64.0368 |
| Object counting | 74.4 | 75.2212 |
| Absolute distance | 57.4 | 57.2422 |
| Object size | 77.8 | 78.1217 |
| Room size | 72.5 | 72.7083 |
| Relative distance | 64.4 | 64.9296 |
| Relative direction | 55.1 | 52.6561 |
| Route planning | 37.6 | 37.6289 |
| Appearance order | 73.5 | 73.7864 |

Raw aggregate results, all 5,130 scored samples, configuration provenance, and [paper_comparison.csv](evidence/table1_qwen_june26_9851_rerun/paper_comparison.csv) are preserved under [table1_qwen_june26_9851_rerun](evidence/table1_qwen_june26_9851_rerun/). Two additional live output records were distributed-padding requests and were excluded from final scoring. A subsequent GPU-7-only spot check reran 24 stratified examples (including eight relative-direction questions) with the same settings: **24/24 raw outputs and cleaned answers were identical** to the four-GPU run. This check found no effect from GPU count on those examples; it does not establish full-split equivalence. See [single-GPU comparison](evidence/table1_qwen_june26_9851_rerun/single_gpu_verification/comparison.json). The full run and logs are under `outputs/qwen_june26_9851/`.

## OOD checkpoint confirmation

The user relayed confirmation from the person who ran the OOD benchmarks that the same consolidated Stage 1+2 and Stage 1+2+3 checkpoints were used:

| OOD experiment | Confirmed local checkpoint ID | Local step |
|---|---|---:|
| Stage 1+2; paper Table 2 BLINK / CV-Bench / NExT-QA | `sn13_sft` | 10000 |
| Corresponding Stage 1+2+3 OOD experiments | `sn13_grpo` | 800 |

These checkpoint identities are accounted for. The paper's Table 2 describes the Stage 1+2 model; no additional GRPO OOD scores or paper rows are inferred from this confirmation. Original OOD evaluation logs/configs have not yet been collected, so their scores are not independently verified from artifacts. The SFT VSI rounding discrepancy below remains recorded. See [checkpoint_identity_confirmations.json](evidence/checkpoint_identity_confirmations.json) for the confirmation source and mappings.

## Matching evidence

Raw result JSONs and available matching per-example JSONLs are preserved under [evidence/](evidence/). [paper_matches.json](evidence/paper_matches.json) records original result locations, full model arguments, sample counts, and unrounded scores. [paper_scores.csv](evidence/paper_scores.csv) provides the task scores in CSV form.

| Experiment | Original result filename | Copied evidence folder |
|---|---|---|
| Table 1 SenseNova / Table 3 final | `20260720_233241_results.json` | [table1_sn13_and_table3_final](evidence/table1_sn13_and_table3_final/) |
| Table 3 SFT (confirmed identity; score discrepancy) | `20260717_061418_results.json` | [table3_sft_candidate](evidence/table3_sft_candidate/) |
| Table 3 VICA-only | `20260727_223552_results.json` | [table3_vica_only](evidence/table3_vica_only/) |
| Table 4 3D only | `20260727_205605_results.json` | [table4_3d_only](evidence/table4_3d_only/) |
| Table 4 box only | `20260729_004258_results.json` | [table4_box_only](evidence/table4_box_only/) |
| Table 4 box + 3D | `20260729_030626_results.json` | [table4_box_and_3d](evidence/table4_box_and_3d/) |
| Table S5 no masking | `20260801_054126_results.json` | [tableS5_no_masking](evidence/tableS5_no_masking/) |
| Table S5 spatial masking | `20260801_060002_results.json` | [tableS5_spatial_masking](evidence/tableS5_spatial_masking/) |
| Table 1 Qwen accepted model, fresh rerun | `20261021_023524_results.json` | [table1_qwen_june26_9851_rerun](evidence/table1_qwen_june26_9851_rerun/) |

All nine result sets report **5,130 effective examples** and all eight VSI tasks. The original eight archived result files were under `/work/old_h200/hunter/covt_test/outputs/easi_lmms3/<run-date>__checkpoint-<step>/`.

For the Table 1 match, the task values below agree with every printed task score after rounding:

| Metric | Paper | Archived evaluation |
|---|---:|---:|
| Average | 68.9 | 68.935723 |
| Object counting | 71.8 | 71.840708 |
| Absolute distance | 59.4 | 59.376499 |
| Object size | 79.2 | 79.233998 |
| Room size | 75.7 | 75.694444 |
| Relative distance | 68.2 | 68.169014 |
| Relative direction | 77.6 | 77.586829 |
| Route planning | 41.8 | 41.752577 |
| Appearance order | 77.8 | 77.831715 |

## Original checkpoint sources and W&B lineage

**Archive cleanup (November 3, 2026):** Seven duplicate source checkpoint directories under `/work/old_h200/hunter/` were removed after fresh SHA-256 verification of all 187 source/copy file pairs against the recorded hashes. The consolidated copies are retained here. Source paths below record historical provenance; see the [cleanup report](evidence/archive_cleanup_20261103.json) for the removed directories and verification evidence.

For `sn13_sft`, the original source is:

`/work/old_h200/hunter/models/sensenova_1_3_st1_newprompt_vsi500/2026-06-23_21-17-15/checkpoint-10000`

For `sn13_stage1`, the original source is `/work/old_h200/moma/checkpoints/checkpoint-11000-universal`.

For `qwen3_sft`, the original source is `/work/old_h200/hunter/shared/clip_qwen3_st1/2026-06-26_16-50-09/checkpoint-9851`.

For the other five SenseNova checkpoints, prepend `/work/old_h200/hunter/covt_test/train/outputs/` to the relative local checkpoint path in the first table. [consolidation_manifest.json](consolidation_manifest.json) contains each absolute source and relative destination.

The GRPO Hydra config explicitly names SFT checkpoint 10000 as `sft_checkpoint`. The selected GRPO weights are from step **800**, rather than the final step 2000. Original run directories named `SN13_SpaceTokens_GRPO` and the shared backup also contain byte-identical `adapter_model.safetensors` at step 800 (SHA-256 `5fdbfcb4eab8ca54f03b0d5ddee060fb6e1a75c47b8fb2691a19479b11538cf6`). The `_eval` copy was selected because the matching Table 1 evaluation explicitly references it.

| Experiment | W&B training run ID |
|---|---|
| SenseNova GRPO | `ywmly68h` |
| VICA-only | `j84eh2ib` |
| 3D only | `8jaoy0kc` |
| Box only | `ibv2414i` |
| Box + 3D | `n592f9s6` |

Selected runs' config YAMLs, final summaries, output logs, and binary `.wandb` histories are under [evidence/wandb/](evidence/wandb/). W&B's final summary often describes the **latest training-time evaluation**, sometimes on a subset with different image settings. The standalone full-split evaluation JSONs establish the paper matches.

## Settings and reproduction details

The matching main-paper standalone evaluations use `covt_qwen2_5_vl`, `prompt_style=covt_cot`, `generation_stage=1`, `torch_dtype=bfloat16`, `max_num_frames=32`, `min_pixels=200704`, `max_pixels=602112`, `attn_implementation=flash_attention_2`, and `interleave_visuals=False`. These values are recorded in each raw result's `config.model_args`; preserve the complete generation/task configuration there when reproducing.

| Checkpoint | Spatial tasks in saved training config | Tokens per task |
|---|---|---|
| SenseNova GRPO / SFT; Qwen SFT | `3d_feature`, `obj_3d_bb` | 8, 8 |
| VICA-only | none (`anchor_model_id: '[]'`) | none |
| Box only | `obj_3d_bb` | 8 |
| 3D only | `3d_feature` | 8 |
| Box + 3D ablation | `obj_3d_bb`, `3d_feature` | 8, 8 |

The VICA-only evaluation's wrapper arguments list unrelated default token IDs, despite the saved training config having no spatial tasks. Both available full evaluations (generation stages 1 and 3) report the same 67.070508 score. Keep this distinction when rebuilding its evaluation command.

Available original `.hydra/` folders remain beside their checkpoint directories so checkpoint metadata discovery can find them. The Stage-1-only export has no original `.hydra/` folder; its provenance is documented above. All selected saved training configs use LoRA rank 16 / alpha 32. Original absolute paths refer to the old server and must be remapped before training or evaluation. Pass the base model explicitly instead of relying on the stale `base_model_name_or_path` in `adapter_config.json`.

The GRPO checkpoint's `non_lora_state_dict.bin` is a **912-byte archive containing an empty dictionary**. This is the original file, not a copy failure. The SFT and ablation checkpoints contain substantial auxiliary weights. The backed-up GRPO loader has auxiliary loading commented out and continues the SFT adapter in its active loading path; that code does not establish the paper's claimed merge-and-new-adapter procedure. Do not silently combine the copied SFT and GRPO adapters or substitute auxiliary weights and assume the recorded scores still apply. The implementation consulted is preserved under [evidence/source_snapshot/](evidence/source_snapshot/).

## Discrepancies and missing matches

- **Table 3 Stage 1 & 2 / Table S3 six frames:** checkpoint 10000 is the explicit GRPO parent and logs 68.667353 on the full split. This would conventionally round to **68.7**, whereas the paper prints **68.6**. Its Stage 1+2 checkpoint identity is now confirmed through the OOD experimenter; the printed VSI score remains a numerical mismatch. The separate 68.416535 box + 3D checkpoint is the shorter Table 4 ablation; it should not replace the SFT parent.
- **Table 2 (BLINK 64.9, CV-Bench 86.9, NExT-QA 80.0):** the OOD experimenter confirmed through the user that Stage 1+2 uses copied `sn13_sft` checkpoint 10000; corresponding Stage 1+2+3 OOD runs use `sn13_grpo` checkpoint 800. Checkpoint identity is accounted for. Original OOD result artifacts/configs are still missing; no new OOD scores are inferred or independently verified.
- **Table 1 Qwen + Space Tokens (64.1):** June 26 checkpoint 9851 is accepted by user instruction as the working paper checkpoint and copied locally as `qwen3_sft`. Its fresh full-split score is 64.0368 (64.0), with task differences and a 24-example single-/four-GPU agreement check documented above. July checkpoint 3000 remains an investigated alternative and was not copied. The original June run also contains checkpoint 9000 and a merged 9851 export; these were not copied. The reason for the paper/rerun score differences is unproven.
- **Pretrained baseline rows (Qwen 59.8, SenseNova 67.6):** model identifiers are specified in the paper, but their exact matching full-split evaluation artifacts were not located. Pretrained base weights were not copied or downloaded. External comparison models are not identified as locally trained checkpoints.
- **Table S3 32-frame and four-frame reasoning:** no exact matching result/configuration pair was located. These are inference variants and do not necessarily require different trained weights. The six-frame SFT candidate is copied as above.
- **Table S4 token replacement:** the text identifies the final SenseNova model, so `sn13_grpo` is the stated model family, but the full-split 68.82 / 68.55 / 68.69 result artifacts were not found. Some archived **100-example** tests happen to score 68.819724; these were deliberately not accepted as matches to the full benchmark.
- **Table S5:** both result files explicitly reference copied GRPO checkpoint 800. Their averages are 63.391573 and 55.254077, matching 63.4 and 55.3, and the masked task vector matches the printed row at the displayed precision. The unmasked relative-direction score is **71.790739** (rounds to 71.8), whereas Table S5 prints **71.9**; its other task scores match. The saved wrapper arguments still say Flash Attention 2 and `trt_mode=zero` for both files, whereas the paper describes eager attention and distinct masking conditions. Thus checkpoint identity and scores are supported, but the exact intervention/backend configuration cannot be reconstructed from these JSON arguments alone. No code/config was changed to conceal this ambiguity.
- **Tables S1–S2:** these are hyperparameter tables, not separate experiments or checkpoints.

The search covered the requested `covt_test` local results and W&B logs, the SFT parent under `hunter/models`, and relevant neighboring `hunter/covt`, `hunter/covt_grpo`, `hunter/shared`, and `moma/covt/train/outputs` backups. The subsequent Stage-1-only search also located and consolidated the parent under `moma/checkpoints`. [evaluation_index.json](evidence/evaluation_index.json) indexes **224** archived evaluations; [wandb_summaries.json](evidence/wandb_summaries.json) indexes **135** available W&B summaries across the two Hunter repositories. No non-VSI benchmark result was present in the indexed evaluations. No remote W&B account was queried.

## Integrity verification

**Completed:** the collection now contains **3,529 verified checkpoint files across 13 directories**, totaling **255.863 GB / 238.291 GiB**. The original archive contains 3,386 files (152 original SenseNova files, 35 Qwen files, and 3,199 Stage-1-only files), verified during their earlier consolidation. The November 12 CVPR additions contribute 143 files in five complete checkpoint directories, all verified against their current sources by SHA-256 and size. See [consolidation_manifest.json](consolidation_manifest.json) for completion status, sizes, steps, and file counts. [SHA256SUMS](SHA256SUMS) and [checkpoint_files.sha256.json](checkpoint_files.sha256.json) record the complete checkpoint-file hashes. Verification compares source and destination file lists, sizes, and SHA-256 digests, including optimizer shards, and checks adapter safetensors header/data extents. Checkpoint pickle payloads were inspected structurally without executing them. One pre-existing structural error was found in the original Table 1 Qwen (`qwen3_sft`, rank 16 / alpha 32) optimizer shard `global_step9850/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt`: ZIP extraction reports `Bad magic number for file header` in both source and destination, whose SHA-256 hashes match. The source file was preserved unchanged. Qwen adapter/auxiliary model weights loaded successfully for the full evaluation, but optimizer-based training resumption must not be assumed to work with this shard. Details are in the Qwen manifest entry under `original_archive_issues`. Archival integrity and numerical reproduction are documented separately; training resumption was not tested.